Quasi-Newton Hamiltonian Monte Carlo

The Hamiltonian Monte Carlo (HMC) method has become significantly popular in recent years. It is the state-of-the-art MCMC sampler due to its more efficient exploration to the parameter space than the standard random-walk based proposal. The key idea behind HMC is that it makes use of first-order gradient information about the target distribution. In this paper, we propose a novel dynamics using second-order geometric information about the desired distribution. The second-order information is estimated by using a quasi-Newton method (say, the BFGS method), so it does not bring heavy computational burden. Moreover, our theoretical analysis guarantees that this dynamics remains the target distribution invariant. As a result, the proposed quasiNewton Hamiltonian Monte Carlo (QNHMC) algorithm traverses the parameter space more efficiently than the standard HMC and produces a less correlated series of samples. Finally, empirical evaluation on simulated data verifies the effectiveness and efficiency of our approach. We also conduct applications of QNHMC in Bayesian logistic regression and online Bayesian matrix factorization problems.